RSE article titled “Leveraging semantic segmentation for photovoltaic plants mapping in optimized energy planning” has been recently published in the Journal Remote Sensing. The article is authored by Giulia Ronchetti and Martina Aiello, Sustainable Development and Energy Sources Department, and Alberto Maldarella, Transmission and Distribution Technologies Department.

The study presents and tests a methodology for the automatic recognition of ground-mounted photovoltaic (PV) plants in Italy, using a semantic segmentation algorithm and satellite images from the Sentinel-2 constellation. The aim of this research is to precisely identify both the locations and the sizes of the plants, estimate their capacity and ensure automatic and regular map updates, therefore providing data that can support energy planning strategies.

For this purpose, a semantic segmentation model, based on a U-Net architecture, has been developed and trained on a Sentinel-2 RGB dataset for the year 2019. The model reached 99% accuracy during the training phase; it has been therefore tested on two distinct cases, including more recent years and different areas of interest. The methodology is based on a multitemporal approach, whereby the semantic segmentation model is applied on a set of images collected throughout the year. The model outputs are then aggregated into a final map that represents the probability of detection of photovoltaic plants.

The proposed methodology has proven to be effective when applied in geographical contexts similar to those of training, but still requires some refinements to ensure a greater applicability in various territorial scenarios. Future developments of this research will be dedicated to improving the generalizability of the model and proposing an automatic tool for the recognition of photovoltaic plants.